AI Copilot & Autonomous Portfolio Risk Officer
Complete operational guide to Firewire AI Copilot: in-chat broker report parsing, automated trade deduplication, REPO exclusion, mathematical risk auditing, and public page intelligence.
AI Copilot is an autonomous analytical assistant based on ReAct architecture. Drag-and-drop any Excel, CSV, or PDF broker report into chat to automatically preview trades, eliminate duplicates, filter technical REPO noise, and confirm ledger updates with one click. Copilot also calculates HHI concentration, tracks dividends, and answers questions on public sharing and comparison pages.
1. Architecture and Operational Principles
Firewire AI Copilot functions as a full-featured autonomous agent following the ReAct (Reasoning + Acting) methodology:
- Context Awareness: dynamically loads positions, open orders, and cash balances across configured accounts.
- MCP Tool Calls: delegates quantitative calculations (HHI, duration, XIRR, dividend projections) to deterministic platform microservices.
- Web Search Toggle (🌐): pulls recent central bank key rate decisions and issuer filings when active.
- GigaChat Engine & BYOK: powered by GigaChat with support for bringing your own Sber Studio API keys for unrestricted access.
2. Interactive Report Import & Smart Deduplication
Upload Excel (.xlsx, .xls), .csv, or .pdf reports directly into the conversation:
- Automated composite matching: checks incoming rows against existing transactions by
ticker + timestamp + operation + execution price + volume, highlighting matches with duplicate badges. - One-Click Action Buttons: «Remove Duplicates» and «Exclude REPO» strip noise and repeated rows instantly.
- Safe Execution: trades are committed only upon explicit «Confirm» button press.
3. AI Copilot on Public Sharing & Comparison Pages
Starting with version 1.75.27, authenticated users can invoke AI Copilot directly on public sharing (/shared/:token) and comparison (/compare) pages:
- Active Page Context Injection: Copilot binds automatically to the token of the displayed portfolio, answering questions strictly regarding its composition.
- Targeted Quick Starters: on shared pages, chips suggest: «What is the risk level?», «Assess diversification», «What is the bond share?». On comparison pages: «Which portfolio is more conservative?», «Compare returns», «Which has higher risk?».
- Public MCP Tools: executes calculations via
get_shared_portfolio_analysisandget_shared_portfolios_comparisonwithout accessing private database tables. - Session & Balance Privacy: chats remain isolated from your private history, and absolute monetary balances stay masked.
4. Data Privacy & PII Scrubbing
- Autonomous PII Sanitization: passport series, tax IDs, and full legal names are stripped prior to external LLM dispatch.
- Isolated Computational Layer: models process anonymized tickers, weights, and durations.
- Deterministic Fallback: in case of LLM timeouts, the platform switches seamlessly to deterministic rule-based algorithms.